Verifying DOPE Model Performance


  • Input images with known dimensions

    • Sample 1080p images for the Ketchup bottle object are provided here

  • Camera intrinsics for the camera used to capture the images

  • Desired image downscale dimensions

    • By default 1920x1080 images are downscaled to 910x512

  • .pth PyTorch weights for an object-specific DOPE model

    • A sample model for the Ketchup bottle object is provided here

  • Object dimensions for the specific object

    • The Ketchup bottle object’s dimensions are included in the source repositories as a default

Run Python DOPE Inference

  1. Check out the DOPE Training repository and install the dependencies as specified in its README:

    git clone
  2. Copy the input image into its own folder:

    mkdir -p /tmp/dope_inference_inputs/ && \
       cp {PATH_TO_IMAGE.jpg} /tmp/dope_inference_inputs
  3. Verify that essential parameters are correctly specified in dope_training/inference/config/config_pose.yaml:

    downscale_height: {DOWNSCALE_HEIGHT}
    dimensions: {
       "{OBJECT_NAME}" : [ 14.860799789428711, 4.3368000984191895, 6.4513998031616211 ],
    thresh_angle: 0.5
    thresh_map: 0.01
    sigma: 3
    thresh_points: 0.0


    Ensure that the dimensions are specified in units of centimeters.

  4. Verify that the camera intrinsics are correctly specified in the projection_matrix entry of dope_training/inference/config/camera_info.yaml:

       rows: 3
       cols: 4
       data: [768.16058349609375, 0, 256, 0, 0, 768.16058349609375, 256, 0, 0, 0, 1, 0]


    Ensure that the camera intrinsics are specified based on the original image dimensions.

  5. Run inference using the DOPE Training repository’s Python inference script:

    cd dope_training/inference && \
       python3 --data /tmp/dope_inference_inputs/ --outf /tmp/dope_inference_outputs/ --object {OBJECT_NAME} --exts jpg --weight {PATH_TO_WEIGHTS.pth}
  6. Verify that the output of the Python inference script exists as .json file:

    ls /tmp/dope_inference_outputs/*/*.json

Run Isaac ROS DOPE Inference

  1. Complete until Run Launch File of the quickstart quickstart.

  2. Instead of step 1 in Run Launch File, run the script with the two additional arguments row and col specifying the desired input image size:

    ros2 run isaac_ros_dope --format onnx \
       --input {PATH_TO_WEIGHTS.pth} --output {PATH_TO_WEIGHTS.onnx} \
  3. Verify that the object dimensions are correctly specified in the dimensions entry of isaac_ros_pose_estimation/isaac_ros_dope/config/dope_config.yaml:

    dimensions: {
       "{OBJECT_NAME}" : [ 14.860799789428711, 4.3368000984191895, 6.4513998031616211 ],

    Note: Ensure that the dimensions are specified in units of centimeters.

  4. Verify that the camera intrinsics are correctly specified in the camera_matrix entry of isaac_ros_pose_estimation/isaac_ros_dope/config/dope_config.yaml:

    camera_matrix: [
      364.16501736,   0.0,          121.36296296,
      0.0,            364.16501736, 121.36296296,
      0.0,            0.0,          1.0


    Ensure that the camera intrinsics are rescaled by multiplying the top two rows of the matrix by a factor of {DOWNSCALED_HEIGHT / INPUT_HEIGHT}.

  5. At step 2 from the Rosbag tab of Run Launch File Section, launch the ROS 2 launch file with two additional arguments network_image_height and network_image_width specifying the desired input image size:

    ros2 launch isaac_ros_dope model_file_path:={PATH_TO_WEIGHTS.onnx} network_image_height:={DOWNSCALED_HEIGHT} network_image_width:={INPUT_WIDTH * DOWNSCALED_HEIGHT / INPUT_HEIGHT}
  6. Open another terminal window and attach to the same container. You should be able to get the poses of the objects in the images through ros2 topic echo:

    cd ~/workspaces/isaac_ros-dev/src/isaac_ros_common && \
    ros2 topic echo /poses
  7. Publish an image using image_publisher:

    ros2 run image_publisher image_publisher_node {PATH_TO_IMAGE} --remap /image_raw:=/image
  8. Save the output logged by ros2 topic echo for comparison later.

    ros2 topic echo /poses >> {PATH_TO_LOG_FILE}

Comparing Outputs between Isaac ROS DOPE and Python DOPE Inference

  1. First, pair the outputs from both DOPE implementations based on the input image used for inference.

  2. For each pair of outputs, run the following steps:

    1. (Optional) Confirm that the Isaac ROS DOPE output detected at least one pose. If your input data contains an image in which the expected result is an empty pose array, then skip this step.

    2. Collect the list of poses found in the Python DOPE Inference’s output .json file.

      Concatenate the location and quaternion_xyzw elements to produce a length-7 pose.

      Multiply the translational components of the pose by a factor of 1/100 to convert the outputs from centimeters to meters.

    3. For each pose in the Isaac ROS DOPE output, run the following steps:

      1. Compare the XYZ position of this pose against those of all poses currently remaining in the Python DOPE Inference output’s list of poses.

      2. Find the closest match based on the L2-norm.

      3. Check that each field of the Isaac ROS DOPE pose and closest-matching Python DOPE Inference pose are equal up to 2 decimal places.

        Note that the sign of the quaternion elements may be flipped, due to the double-cover nature of quaternions.

      4. Remove the consumed Python DOPE Inference pose from the list of poses, so that it is only matched to one Isaac ROS DOPE pose.

  3. If all checks for all poses for all output pairs pass, the verification was successful.

NVIDIA has run this testing process in an automated fashion using the Ketchup suite of example data.